« VWURQJ VXSHUYLVLRQ - Inria · e\ /lqgd 6plwk dqg 0lfkdho *dvvhu $ IRUP RI XQVXSHUYLVHG OHDUQLQJ...

$Page 1: « VWURQJ VXSHUYLVLRQ - Inria · e\ /lqgd 6plwk dqg 0lfkdho *dvvhu $ IRUP RI XQVXSHUYLVHG OHDUQLQJ ZKHUH WKH GDWD SURYLGHV WKH VXSHUYLVLRQ ,Q JHQHUDO ZLWKKROG VRPH SDUW RI WKH GDWD$
Self-Supervised Learning

Andrew Zisserman

Slides from: Carl Doersch, Ishan Misra, Andrew Owens, Carl Vondrick, Richard Zhang

1000 categories

• Training: 1000 images for each category

• Testing: 100k images

The ImageNet Challenge Story …

The ImageNet Challenge Story … strong supervision

The ImageNet Challenge Story … outcomes

Strong supervision:

• Features from networks trained on ImageNet can be used for other visual tasks, e.g. detection, segmentation, action recognition, fine grained visual classification

• To some extent, any visual task can be solved now by:1. Construct a large-scale dataset labelled for that task

2. Specify a training loss and neural network architecture

3. Train the network and deploy

• Are there alternatives to strong supervision for training? Self-Supervised learning ….

1. Expense of producing a new dataset for each new task

2. Some areas are supervision-starved, e.g. medical data, where it is hard to obtain annotation

3. Untapped/availability of vast numbers of unlabelled images/videos

– Facebook: one billion images uploaded per day

– 300 hours of video are uploaded to YouTube every minute

4. How infants may learn …

Why Self-Supervision?

Self-Supervised Learning

The Scientist in the Crib: What Early Learning Tells Us About the Mind by Alison Gopnik, Andrew N. Meltzoff and Patricia K. Kuhl

The Development of Embodied Cognition: Six Lessons from Babiesby Linda Smith and Michael Gasser

• A form of unsupervised learning where the data provides the supervision

• In general, withhold some part of the data, and task the network with predicting it

• The task defines a proxy loss, and the network is forced to learn what we really care about, e.g. a semantic representation, in order to solve it

What is Self-Supervision?

Randomly Sample PatchSample Second Patch

CNN CNN

Classifier

8 possible locations

Example: relative positioning

Train network to predict relative position of two regions in the same image

Unsupervised visual representation learning by context prediction, Carl Doersch, Abhinav Gupta, Alexei A. Efros, ICCV 2015

A B

Example: relative positioning


Semantics from a non-semantic task


What is learned?

Relative-positioningInput Random Initialization ImageNet AlexNet

CNN CNN

Classifier

Self-supervised learning in three parts:

1. from images

2. from videos

3. from videos with sound

Outline

Part I

Self-Supervised Learning from Images

Randomly Sample PatchSample Second Patch

CNN CNN

Classifier

8 possible locations

Recap: relative positioning

Train network to predict relative position of two regions in the same image


Evaluation: PASCAL VOC Detection

• 20 object classes (car, bicycle, person, horse …)

• Predict the bounding boxes of all objects of a given class in an image (if any)

Dog Horse Motorbike Person

Pre-train on relative-position task, w/o labels

[Girshick et al. 2014]


• Pre-train CNN using self-supervision (no labels)

• Train CNN for detection in R-CNN object category detection pipeline

R-CNN

45.6%

No PretrainingRelative positioning

ImageNet Labels

51.1%56.8%

Aver

age

Prec

isio

n


Avoiding Trivial Shortcuts

Include a gap

Jitter the patch locations

Position in Image

A Not-So “Trivial” Shortcut

Chromatic Aberration

Position in Image

A Not-So “Trivial” Shortcut

Solution?

Only use one of the colour channels

abL

Concatenate (L,ab)Grayscale image: L channel

“Free” supervisory

signal

Image example II: colourizationTrain network to predict pixel colour from a monochrome input

Image example II: colourizationTrain network to predict pixel colour from a monochrome input

Colorful Image Colorization, Zhang et al., ECCV 2016

- Exemplar Networks (Dosovitskiy et al., 2014)

- Perturb/distort image patches, e.g. by cropping and affine transformations

- Train to classify these exemplars as same class

Image example III: exemplar networks

Egomotion

Agrawal et al. ICCV 2015 Jayaraman et al. ICCV 2015Isola et al. ICLR Workshop 2016.

Context

Noroozi et al 2016 Pathak et al. CVPR 2016

Hinton & Salakhutdinov.Science 2006.

Zhang et al. CVPR 2017

Dosovitskiy et al., NIPS 2014

Split-brain auto-encoders

Co-Occurrence

Exemplar networksAutoencoders Denoising Autoencoders

Vincent et al. ICML 2008.

Multi-Task Self-Supervised Learning

Self-supervision task ImageNet Classification

top-5 accuracy

PASCAL VOC Detection

mAP

Rel. Pos 59.21 66.75

Colour 62.48 65.47

Exemplar 53.08 60.94

Rel. Pos + colour 66.64 68.75

Rel. Pos + Exemplar 65.24 69.44

Rel. Pos + colour + Exemplar 68.65 69.48

ImageNet labels 85.10 74.17

Procedure:

• ImageNet-frozen: self-supervised training, network fixed, classifier trained on features

• PASCAL: self-supervised pre-training, then train Faster-RCNN

• ImageNet labels: strong supervision

NB: all methods re-implemented on same backbone network (ResNet-101)

Multi-task self-supervised visual learning, C Doersch, A Zisserman, ICCV 2017

Multi-Task Self-Supervised Learning

Multi-task self-supervised visual learning, C Doersch, A Zisserman, ICCV 2017

Self-supervision task ImageNet Classification

top-5 accuracy

PASCAL VOC Detection

mAP

Rel. Pos 59.21 66.75

Colour 62.48 65.47

Exemplar 53.08 60.94

Rel. Pos + colour 66.64 68.75

Rel. Pos + Exemplar 65.24 69.44

Rel. Pos + colour + Exemplar 68.65 69.48

ImageNet labels 85.10 74.17Procedure:

• ImageNet-frozen: self-supervised training, network fixed, classifier trained on features

• PASCAL: self-supervised pre-training, then train Faster-RCNN

• ImageNet labels: strong supervision

Findings:

• Deeper network improves performance (ResNet vs AlexNet)

• Colour and Rel-Pos superior to Exemplar

• Gap between self-supervision and strong supervision closing

Image Transformations – 2018

Unsupervised representation learning by predicting image rotations, Spyros Gidaris, Praveer Singh, Nikos Komodakis, ICLR 2018

Which image has the correct rotation?






• Uses AlexNet

• Closes gap between ImageNet and self-supervision

PASCAL VOC Detection mAP

Random 43.4

Rel. Pos. 51.1

Colour 46.9

Rotation 54.4

ImageNet Labels 56.8


Summary Point

• Self-Supervision:

– A form of unsupervised learning where the data provides the supervision

– In general, withhold some information about the data, and task the network with predicting it

– The task defines a proxy loss, and the network is forced to learn what we really care about, e.g. a semantic representation, in order to solve it

• Many self-supervised tasks for images

• Often complementary, and combining improves performance

• Closing gap with strong supervision from ImageNet label training

– ImageNet image classification, PASCAL VOC detection

• Deeper networks improve performance

Part II

Self-Supervised Learning from Videos

Video

A temporal sequence of frames

What can we use to define a proxy loss?

• Nearby (in time) frames are strongly correlated, further away may not be

• Temporal order of the frames

• Motion of objects (via optical flow)

• …

Three example tasks:

– Video sequence order

– Video direction

– Video tracking

Outline

Temporal structure in videos

Time

“Sequence” of data

Shuffle and Learn: Unsupervised Learning using Temporal Order Verification

Ishan Misra, C. Lawrence Zitnick and Martial Hebert ECCV 2016

Slide credit: Ishan Misra

Sequential Verification

• Is this a valid sequence?

Sun and Giles, 2001; Sun et al., 2001; Cleermans 1993; Reber 1989Arrow of Time - Pickup et al., 2014


Orig

inal

vid

eo


Orig

inal

vid

eo

Temporally Correct orderTemporally Correct order


Orig

inal

vid

eo

Temporally Correct orderTemporally Correct order

Temporally Incorrect order Slide credit: Ishan Misra

Geometric View

Given a start and an end, can this point lie in between?

Images

Shuffle and Learn – I. Misra, L. Zitnick, M. Hebert – ECCV 2016 Slide credit: Ishan Misra

Dataset: UCF-101 Action Recognition

UCF101 - Soomro et al., 2012

Positive Tuples Negative Tuples

~900k tuples from UCF-101 dataset (Soomro et al., 2012)Slide credit: Ishan Misra

Orig

inal

vid

eo

Temporally Correct orderInformative training tuples

Frame Motion

Time

High motion window


conc

aten

atio

n

Input Tuple

fc8

Correct/IncorrectTuple

class

ifica

tion

Cross Entropy Loss


Query ImageNet Shuffle & Learn Random

Nearest Neighbors of Query Frame (fc7 features)


Finetuning setup

conc

aten

atio

n

Correct/IncorrectTuple

Input Tuple

clas

sifica

tion

Test -> FinetuneSelf-supervised Pre-train

Action Labels


Results: Finetune on Action Recognition

Dataset Initialization Mean ClassificationAccuracy

UCF101 Random 38.6Shuffle & Learn 50.2

ImageNet pre-trained 67.1


Setup from - Simonyan & Zisserman, 2014

What does the network learn?

Given a start and an end, can this point lie in between?

Images

Shuffle and Learn – I. Misra, L. Zitnick, M. Hebert – ECCV 2016 Slide credit: Ishan Misra

Human Pose Estimation

• Keypoint estimation using FLIC and MPII Datasets


FLIC Dataset MPII DatasetInitialization Mean PCK AUC PCK Mean

[email protected]

[email protected] & Learn 84.9 49.6 87.7 47.6

ImageNet pre-train 85.8 51.3 85.1 47.2

FLIC - Sapp & Taskar, 2013MPII - Andriluka et al., 2014Setup fom – Toshev et al., 2013


Human Pose Estimation

• Keypoint estimation using FLIC and MPII Datasets

More temporal structure in videos

Self-Supervised Video Representation Learning With Odd-One-Out Networks

Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould, ICCV 2017

More temporal structure in videos

Self-Supervised Video Representation Learning With Odd-One-Out Networks

Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould, ICCV 2017

Initialization Mean Classification

AccuracyRandom 38.6

Shuffle and Learn 50.2

Odd-One-Out 60.3

ImageNet pre-trained

67.1

• Important to select informative data in training– Hard negatives and positives

– Otherwise, most data is too easy or has no information and the network will not learn

– Often use heuristics for this, e.g. motion energy

• Consider how the network can possibly solve the task (without cheating)– This determines what it must learn, e.g. human keypoints in `shuffle and learn’

• Choose the proxy task to encourage learning the features of interest

Summary: lessons so far

Self-Supervision using the Arrow of Time

Donglai Wei, Joseph Lim, Bill Freeman, Andrew Zisserman CVPR 2018

Learning the arrow of time

Supervision:

Positive training samples: video clips playing forwards

Negative training samples: video clips playing backwards

Task: predict if video playing forwards or backwards

Strong cues

Semantic, face motion direction, ordering


Strong cues

`Simple’ physics:• gravity• entropy• friction• causality


Weak or no cues

Symmetric in time, constant motion, repetitions


Temporal Class-Activation Map Network

input motion

1 … 10

forwardsor

backwards?

11 … 20

T-CAM Model:

Input: optical flow in two chunks

Final layer: global average pooling to allow class activation map (CAM)

The inevitable cheating …

Dataset: UCF-101 actions

Train/Test: 70%/30%

AoT Test accuracy: 98%

Chance accuracy: 50%

Cautionary tale: Chromatic aberration used as shortcut in Doersch C, Gupta A, Efros AA, Unsupervised visual representation learning by context prediction. ICCV 2015


Cue I: black framing

when black stripe signals are zeroed-out,test accuracy drops ~10%

0 5

timeh

eig

ht

black stripes are not “purely black”

46% of videos have black framing


cluster A(camera zoom-in)

cluster B(camera tilt-down)

K-means clustering on test clips with top scores

Cue II: cinematic conventions

73% of videos have camera motion

when camera motion is stabilized, test accuracy drops ~10%

original camera stabilized

(black stripe removed)

Stabilize to remove camera motion/zoom

Datasets and Performance

Flickr 150K shots

• Obtained from 1.74M shots used in Thomee et al (2016) &Vondrick et al (2016), after black stripe removal and stabilization

• Split 70:30 for train:test

Model accuracy on test set: 81%

Human accuracy on test set: 81%

Chance: 50%


“Semantic” motionsin

put

vid

eop

redi

ctio

n he

atm

ap

backward

forward

Evaluation: Action Classification

Procedure:

• Pre-train network

• Fine tune & test network on UCF101 human action classification benchmark

• * = Wang et al, Temporal Segment Networks, 2016 (also VGG-16 and flow, pre-trained on ImageNet)

Pre-train Performance

T-CAM on AoT on Flickr 150k shots 84.1

T-CAM on AoT on UCF-101 86.3

Flow network on ImageNet* 85.7


Tracking Emerges by Colorizing Videos

Vondrick, Shrivastava, Fathi, Guadarrama, Murphy, ECCV 2018

Color is mostly temporally coherent

Temporal Coherence of Color

RGB

ColorChannels

QuantizedColor

Self-supervised Tracking

Vondrick, Shrivastava, Fathi, Guadarrama, Murphy. ECCV 2018.

Reference Frame Gray-scale Video

Task: given a color video … Colorize all frames of a gray scale version using a reference frame

What color is this?

Where to copy color from?

Semantic correspondence

Input Frame


Target Colors

Input FrameReference Frame

Reference Colors

Colorize by Pointing


Target Colors


Reference Colors


Target Colors


Reference Colors


Target Colors


Reference Colors


Target Colors


Reference Colors


Video Colorization

Ground TruthReference Frame Gray-scale Video Predicted Color

Train: Kinetics

Evaluate: DAVIS


Visualizing Embeddings

Project embedding to 3 dimensions and visualize as RGB

Train: Kinetics

Evaluate: DAVIS


Tracking Emerges!Input FrameReference Frame

Predicted MaskReference Mask


Only the first frame is given. Colors indicate different instances.

Segment Tracking Results


Pose Tracking Results

Only the skeleton in the first frame is given.


Part III

Self-Supervised Learning from Videos with Sound

Audio-Visual Co-supervision

Sound and frames are:

• Semantically consistent

• Synchronized


Objective: use vision and sound to learn from each other

• Two types of proxy task:

1. Predict audio-visual correspondence

2. Predict audio-visual synchronization


Train a network to predict if image and audio clip correspond

Correspond?

“Objects that Sound”, Arandjelović and Zisserman, ICCV 2017 & ECCV 2018

Audio-Visual Correspondence

drum

guitar

drum

guitar

positive


drum

guitar

positive


drum

guitar

negative


Correspond?yes/no

Contrastive loss based on distance

between vectors

Distance between audio and visual vectors:

• Small: AV from the same place in a video (Positives)

• Large: AV from different videos (Negatives)

Train network from scratch

visual subnetwork

audio subnetwork

single frame

1 s

Audio-Visual Embedding (AVE-Net)

Overview

What can be learnt by watching and listening to videos?

• Good representations– Visual features

– Audio features

• Intra- and cross-modal retrieval– Aligned audio and visual embeddings

• “What is making the sound?”– Learn to localize objects that sound


• Andrew Owens ….– Owens, A., Jiajun, W., McDermott, J., Freeman, W., Torralba, A.: Ambient sound provides

supervision for visual learning. ECCV 2016

– Owens, A., Isola, P., McDermott, J., Torralba, A., Adelson, E., Freeman,W.: Visually indicated sounds. CVPR 2016

• Other MIT work:– Aytar, Y., Vondrick, C., Torralba, A.: SoundNet: Learning sound representations from

unlabeled video. NIPS 2016

• From the past:

– Kidron, E., Schechner, Y.Y., Elad, M.: Pixels that sound. CVPR 2005

– De Sa, V.: Learning classification from unlabelled data, NIPS 1994

Background: Audio-Visual

Dataset

- AudioSet (from YouTube), has labels

- 200k x 10s clips

- use musical instruments classes

- Correspondence accuracy on test set: 82% (chance: 50%)


Use audio and visual features



– Audio features



correspond? yes/no

visual subnetwork

audio subnetwork

single frame

1 s


Sound classification• ESC-50 dataset

– Environmental sound classification

– Use the net to extract features

– Train linear SVM

Results: Audio features


ImageNet classification• Standard evaluation procedure for unsupervised / self-supervised setting

– Use the net to extract visual features

– Linear classification on ImageNet

• On par with state-of-the-art self-supervised approaches

• The only method whose features haven’t seen ImageNet images

– Probably never seen ‘Tibetan terrier’

– Video frames are quite different from images

Results: Vision features




– Audio features



correspond? yes/no

visual subnetwork

audio subnetwork

single frame

1 s


Query on image, retrieve audio

Search in 200k video clips of AudioSet

Query frame

Top 10 ranked audio clips





– Audio features



correspond? yes/no

visual subnetwork

audio subnetwork

single frame

1 s


Objects that Sound

Audio embedding

Apply Visual ConvNetconvolutionally

AVE-Net

Visual embedding

Single audio representation

128-D

14x14 spatial grid of 128-D visual

representations

AVOL-Net

Multiple instance learning


Localizing objects with sound

Input: audio and video frame

Output: localization heatmap on frame

What would make this sound?

Note, no video (motion) information is used


To embed or not to embed?

Audio embeddingVisual embedding

Concatenation Embedding

Features available

Cross-modal alignment in embedding

Specialize to talking heads …

Objective: use faces and voice to learn from each other




Specialize to talking heads …

Objective: use faces and voice to learn from each other




Lip-sync problem on TV

Face-Speech Synchronization

• Positive samples: in sync

• Negative samples: out of sync (introduce temporal offset)

Chung, Zisserman (2016) “Out of time: Automatic lip sync in the wild”

Sequence-sequence face-speech network

• The network is trained with contrastive loss to:

– Minimise distance between positive pairs

– Maximise distance between negative pairs

Contrastive loss


Face-Speech Synchronization


Averaged sliding windows

The predicted offset value is >99% accurate, averaged over 100 frames.

In-sync Off-sync Non-speaker

Dis

tan

ce

Offset Offset Offset

Application: Lip Synchronization

Application: Active speaker detection

Blue: speaker Red: non-speaker

Face-Speech Synchronization - summary

The network can be used for:

– Audio-to-video synchronisation

– Active speaker detection

– Voice-over rejection

– Visual features for lip reading

Audio-Visual Synchronization

Self-supervised Training

Audio-Visual Scene Analysis with Self-Supervised Multisensory Features, Andrew Owens, Alyosha Efros, 2018

Misaligned Audio

Shifted audio track


Visualizing the location of sound sources

3D class activation map

3D Convolution

3D Convolution

3D Convolution

3D Convolution 1D Convolution

1D Convolution

1D Convolution



Summary: Audio-Visual Co-supervision

Objective: use vision and sound to learn from each other


1. Predict audio-visual correspondence -> semantics

2. Predict audio-visual synchronization -> attention

• Lessons are applicable to any two related sequences, e.g. stereo video, RGB/D video streams, visual/infrared cameras …

• Self-Supervised Learning from images/video– Enables learning without explicit supervision

– Learns visual representations – on par with ImageNet training

• Self-Supervised Learning from videos with sound– Intra- and cross-modal retrieval

– Learn to localize sounds

– Tasks not just a proxy, e.g. synchronization, attention, applicable directly

• Applicable to other domains with paired signals, e.g. – face and voice

– Infrared/visible

– RGB/D

– Stereo streams …

Summary

Date post:	05-Oct-2020
Category:	Documents
Upload:	others
View:	0 times
Download:	0 times

« VWURQJ VXSHUYLVLRQ - Inria · e\ /lqgd 6plwk dqg 0lfkdho *dvvhu $ IRUP RI XQVXSHUYLVHG OHDUQLQJ...

Documents